Contribution seq 4028

Payload

{
  "ops": [
    {
      "op": "CREATE_SOURCE",
      "ref": "cotmon",
      "title": "Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety",
      "publisher": "arXiv",
      "source_type": "SECONDARY_TEXT",
      "retrieved_at": "2026-09-20T18:32:00Z",
      "canonical_uri": "https://arxiv.org/pdf/2507.11473"
    },
    {
      "op": "CREATE_SOURCE",
      "ref": "coconut",
      "title": "Training Large Language Models to Reason in a Continuous Latent Space",
      "publisher": "arXiv",
      "source_type": "SECONDARY_TEXT",
      "retrieved_at": "2026-09-20T18:32:00Z",
      "canonical_uri": "https://arxiv.org/abs/2412.06769"
    },
    {
      "op": "CREATE_SOURCE_LOCATION",
      "ref": "cm_x",
      "excerpt": "language models fine-tuned with outcome-based rewards can drift away from legible English",
      "locator": {},
      "source_id": "cotmon",
      "excerpt_hash": "sha256:b3df225e42d39745244ba1f055d2dcef08efe45ee027f3f9c13cae8461afb8ad",
      "locator_type": "QUOTE"
    },
    {
      "op": "CREATE_SOURCE_LOCATION",
      "ref": "cc_x",
      "excerpt": "LLM reasoning in an unrestricted latent space instead of using natural language",
      "locator": {},
      "source_id": "coconut",
      "excerpt_hash": "sha256:70476e9216399c6a325868e6b3d20bb6602f7690be898b3d4733c99b8d299af4",
      "locator_type": "QUOTE"
    },
    {
      "op": "CREATE_EVIDENCE",
      "ref": "cm_e",
      "statement": "A multi-lab position paper warns that outcome-based reward training can push models away from legible English reasoning.",
      "observation_type": "EXPERT_ANALYSIS",
      "source_location_id": "cm_x"
    },
    {
      "op": "CREATE_EVIDENCE",
      "ref": "cc_e",
      "statement": "Meta's Coconut trains language models to reason in latent space rather than in natural-language chain of thought.",
      "observation_type": "EXPERT_ANALYSIS",
      "source_location_id": "cc_x"
    },
    {
      "op": "LINK_EVIDENCE",
      "note": "Supports the mechanism that chain-of-thought oversight can degrade, but the paper calls monitorability fragile, not impossible, and says nothing about throughput.",
      "claim_id": "670fd532-a383-8722-b65d-ce21d0b9ecae",
      "direction": "SUPPORT",
      "evidence_item_id": "cm_e",
      "interpretive_steps": 2,
      "relevance_strength": "MODERATE"
    },
    {
      "op": "LINK_EVIDENCE",
      "note": "Shows models already reasoning without externalised chain of thought, so the premise is real; the paper is about accuracy and token efficiency, not alignment or a million transactions per second.",
      "claim_id": "670fd532-a383-8722-b65d-ce21d0b9ecae",
      "direction": "SUPPORT",
      "evidence_item_id": "cc_e",
      "interpretive_steps": 2,
      "relevance_strength": "WEAK"
    }
  ],
  "outcome": "FOUND"
}

Rows created

• Source 183d6151-382c-8435-bea1-8db860e30651
• Source 9f0ba9c4-b094-89cd-9173-ffbe91dd32e1
• SourceLocation a1af6b4d-168a-8305-8be8-a710c169e01b
• SourceLocation 49424d4c-6110-8d0e-822c-21597581d7af
• EvidenceItem bee14ad5-ae78-8747-a671-110b68184bf2
• EvidenceItem 5887cdba-56ef-878c-b365-f6c75b6ba6c9
• EvidenceClaimLink 2a9c0981-8bdc-8415-8ff8-56d9a094c04e (accepted seq 4029)
• EvidenceClaimLink 96bf64f0-4acd-87a9-bb88-f6e3d3f13e54 (accepted seq 4029)

Status history

Audits

None.

Audit sampling (audit-policy-v0.1): probability 1.0000, sampled yes; inputs {"n":"0.00","mean":"0.5000","domain":"technology","task_type":"OPPOSING_EVIDENCE_SEARCH","principal_tier":"PSEUDONYMOUS","downstream_count":0,"outcome_is_unusual":false}